A genome-wide design and an empirical partially Bayes approach to increase the power of Mendelian randomization, with application to the effect of blood lipids on cardiovascula
share
Mendelian randomization (MR) is an instrumental variable method of estimating the causal effect of risk exposures in epidemiology, where genetic variants are used as instruments. With the increasing availability of large-scale genome-wide association studies, it is now possible to greatly improve the power of MR by using genetic variants that are only weakly relevant. We consider how to increase the efficiency of Mendelian randomization by a genome-wide design where more than a thousand genetic instruments are used. An empirical partially Bayes estimator is proposed, where weaker instruments are shrunken more heavily and thus brings less variation to the MR estimate. This is generally more efficient than the profile-likelihood-based estimator which gives no shrinkage to weak instruments. We apply our method to estimate the causal effect of blood lipids on cardiovascular diseases. We find high-density lipoprotein cholesterol (HDL-c) has a significantly protective effect on heart diseases, while previous MR studies reported null findings.
READ FULL TEXT
